News Nvidia CEO Says Intel's Test Chip Results For Next-Gen Process Are Good

[bit_user]

Ok, in that case prices would stay the same which isn't the case at nvidia.

I don't understand how you reached that conclusion, since their current GPUs have significantly more transistors than the last gen. I think I explained quite adequately, in my next post.

I lack the cost data on Samsung's node to project how much more the RTX 4090 should cost than the RTX 3090, but we can reasonably assume Samsung 8 nm was cheaper per Tr than TSMC N7, or else why would Nvidia have switched? Therefore, the price differential should be even greater than the 232% I estimated for TSMC N7.

 

[thestryker]

I don't understand how you reached that conclusion, since their current GPUs have significantly more transistors than the last gen. I think I explained quite adequately, in my next post.

I lack the cost data on Samsung's node to project how much more the RTX 4090 should cost than the RTX 3090, but we can reasonably assume Samsung 8 nm was cheaper per Tr than TSMC N7, or else why would Nvidia have switched? Therefore, the price differential should be even greater than the 232% I estimated for TSMC N7.

Since transistor counts are relatively meaningless when it comes to chip cost the only thing that matters is cost per wafer and number of good chips you get out of said wafer. While nobody knows the exact cost of 4N the general consensus has been that it's roughly twice the cost per wafer as nvidia was paying for Samsung's "8nm". FWIW I believe the density difference between N7 and "8nm" ended up being about 15% with RNDA2 vs Ampere though a lot more goes into that than just chip sizes. TSMC was also raising prices during that round which I'm sure Samsung was happy to undercut to get the business.

An exercise in pure conjecture:
If we say $20000 for N4 and $10000 for 8nm and a 75% yield:

AD102 vs GA102:
You end up with 66 good chips for each so that's ~$303 for AD102 and ~$151 for GA102.

Now for AD103 vs GA102 (4080 vs 3080):
AD103 is 111 and GA102 is again 66 so that's ~$180 for AD102 and again ~$151 for GA102

AD104 vs GA104 (4070/4070ti vs 3060ti/3070/3070ti):
AD104 is 145 and GA104 is 111 so that's ~$138 for AD104 and ~$90 for GA104

This is why the RTX 4090, relatively speaking, comes across as a decent deal to me. I wouldn't get one because I'm not interested in spending that much money on a single part, but relatively speaking the $100 price jump seems perfectly reasonable over the RTX 3090. The other chips in the 40 series start getting smaller, come with a narrower memory bus and lower power consumption. So you end up with a card that should be cheaper to produce and chips which aren't much more expensive yet the prices have gone up by $100 3070 to 4070, $200 3070ti to 4070ti, and $500 3080 to 4080.

Now obviously it's not that simple as we just don't have enough concrete numbers, but that shouldn't be far off reality in terms of relative price differences. There's clearly abuse of market position and an urge to sate shareholders without a crypto boom (though I'm sure nvidia is thrilled about the AI boom) margins wise.

 

[bit_user]

This is why the RTX 4090, relatively speaking, comes across as a decent deal to me. I wouldn't get one because I'm not interested in spending that much money on a single part, but relatively speaking the $100 price jump seems perfectly reasonable over the RTX 3090. The other chips in the 40 series start getting smaller, come with a narrower memory bus and lower power consumption. So you end up with a card that should be cheaper to produce and chips which aren't much more expensive yet the prices have gone up by $100 3070 to 4070, $200 3070ti to 4070ti, and $500 3080 to 4080.

If you simply look at cost per mm^2 and cost per GB, the 4070 Ti's MSRP is actually in line with the 4090's! Among the first 3 cards to launch, it's really the 4080 that's the outlier. By my estimates, the 4080 should sell for about $1050, not $1200.

There's clearly abuse of market position and an urge to sate shareholders without a crypto boom (though I'm sure nvidia is thrilled about the AI boom) margins wise.

I think the crypto boom showed Nvidia what people could & would pay for a gaming card. As a result, they adjusted their price targets and then upgraded the hardware spec to fit. When the market collapsed, that left them relatively unable to respond with price cuts, because the costs were so much higher.

 

[thestryker]

If you simply look at cost per mm^2 and cost per GB, the 4070 Ti's MSRP is actually in line with the 4090's! Among the first 3 cards to launch, it's really the 4080 that's the outlier. By my estimates, the 4080 should sell for about $1050, not $1200.

Cost per mm^2 is not really a great metric to use seeing as multiple SKUs use the same die (ex: 4070 and 4070ti use AD104). Cost per GB is also nebulous seeing as they're using fewer memory chips per tier than the 30 series counterpart which impacts board cost. It's also extremely unlikely that GDDR6X is costing nvidia per GB what it did when the 30 series came out.

I think the crypto boom showed Nvidia what people could & would pay for a gaming card. As a result, they adjusted their price targets and then upgraded the hardware spec to fit. When the market collapsed, that left them relatively unable to respond with price cuts, because the costs were so much higher.

Their costs are not much higher than they were with the 30 series which is the point. The 4060ti for example is only PCIe x8 which cuts down on board cost further not to mention the die size is less than half that of the 3060ti. They knew the crypto boom was just that well before the SKUs and pricing for the 40 series was decided upon. The lesson they seem to have taken from the crypto boom was certainly people would pay more, but that doesn't mean that their costs went up accordingly.

 

[bit_user]

Cost per mm^2 is not really a great metric to use seeing as multiple SKUs use the same die (ex: 4070 and 4070ti use AD104).

If we assume that the 4070 Ti and 4090 both enable about the same % of their respective dies, then it's a fine metric.

Cost per GB is also nebulous seeing as they're using fewer memory chips per tier than the 30 series counterpart which impacts board cost.

The number of memory chips and data bus width scales linearly between the three SKUs I mentioned.

I think you're assuming that I'm talking about 4000-series vs. 3000-series, but I'm not. I'm only comparing 3 products: RTX 4090, RTX 4080, RTX 4070 Ti. What I'm saying is that if you plot the 3 of them in terms of GB/$ or mm^2/$, then the 4080 is the outlier. It's overpriced by about $150.

Their costs are not much higher than they were with the 30 series which is the point.

I thought we were in agreement that the 4000-series GPU dies appear to cost considerably more. If you don't see that, then you should review the data.

The 4060ti for example is only PCIe x8 which cuts down on board cost further not to mention the die size is less than half that of the 3060ti.

I'm skeptical just how much difference it makes, at either the die or board level. Still, even a couple $ is probably worth it.

As for why they decided to x8 now, I think we're seeing the impact of greater PCIe 4.0 penetration. I think they're betting that a card of that tier just doesn't need full x16. It won't make much difference, on a PCIe 4.0 motherboard. More, on a PCIe 3.0 board, but still not a deal-breaker.

I guess games tend to stress PCIe bandwidth pretty hard, when you're using settings that put too much pressure on GPU memory. At least, that's what I think we saw with the RX 6500XT. With settings turned up well past the point that its meager 4 GB filled up, performance would be greatly affected based on whether it was in a PCIe 3.0 or 4.0 system.

They knew the crypto boom was just that well before the SKUs and pricing for the 40 series was decided upon.

That's impossible. I mean, everyone knew the risk of the crypto boom ending, but the GPU dies' specs were finalized probably 2 years before launch. You have to know what your price floor is and the approximate wafer costs, at that point in the development process. You have to build the right product for the market you expect to launch it into.

Once you've committed to the chip specs, about the only choice you have left is to cancel any that wouldn't be commercially viable, and that means all the development costs get flushed down the toilet and you get zero return-on-investment. So, it's critical to get the pricing right, years before launch.

The lesson they seem to have taken from the crypto boom was certainly people would pay more,

That's what I'm saying.

but that doesn't mean that their costs went up accordingly.

You just showed their GPU price at least doubled. Nvidia needs to make its profits, so they probably do about a 100% markup. Not only that, but a larger, hotter, high-power board requires more expensive VRMs and cooling solution. On top of that, their board partners have costs and need to make a profit.

The design costs for these chips is starting to starting to stretch into the $Billions. Also, people tend to be dismissive of software costs. Something like 90% of Nvidia's engineers are software. They need to get paid, too! Traditionally, you can ignore these NRE costs, but I think they're actually big enough that I think you simply cannot just look at chip fabrication costs, to establish a cost basis for them.

 

[thestryker]

If we assume that the 4070 Ti and 4090 both enable about the same % of their respective dies, then it's a fine metric.

The 4070 uses the same die as the 4070ti and the cost to manufacture is die size so it doesn't matter how much is "enabled". The yield might be slightly higher, but there's a $200 MSRP difference between the two.

I think you're assuming that I'm talking about 4000-series vs. 3000-series, but I'm not. I'm only comparing 3 products: RTX 4090, RTX 4080, RTX 4070 Ti. What I'm saying is that if you plot the 3 of them in terms of GB/$ or mm^2/$, then the 4080 is the outlier. It's overpriced by about $150.

Yeah I was, I get what you mean now.

I thought we were in agreement that the 4000-series GPU dies appear to cost considerably more. If you don't see that, then you should review the data.

They don't though, because they're largely a lot smaller than their 30 series counterparts so nvidia gets more per wafer. This is why the 4090 is such an anomaly in that it's almost the same size as the 3090, but the price increase wasn't very large.

That's impossible. I mean, everyone knew the risk of the crypto boom ending, but the GPU dies' specs were finalized probably 2 years before launch. You have to know what your price floor is and the approximate wafer costs, at that point in the development process. You have to build the right product for the market you expect to launch it into.

Once you've committed to the chip specs, about the only choice you have left is to cancel any that wouldn't be commercially viable, and that means all the development costs get flushed down the toilet and you get zero return-on-investment. So, it's critical to get the pricing right, years before launch.

They absolutely can, and did. You're mistaking what I said thinking I'm talking about the chip design when I'm definitely not. I'm talking about the SKUs and subsequent pricing which we all watched them change in public after backlash (12GB 4080 turning into 4070ti). It's pretty obvious that the 4060ti-4070ti are all 1 tier off what they really should have been when comparing performance level versus prior generation uplifts.

You just showed their GPU price at least doubled. Nvidia needs to make its profits, so they probably do about a 100% markup.

The GPU price didn't double the wafer price did which means the prices went up, but how much is dependent entirely on how many chips per wafer they're getting. The 4060ti will forever be my punching bag example here because it's less than half the size of the 3060ti.

Not only that, but a larger, hotter, high-power board requires more expensive VRMs and cooling solution.

The 4090 has a higher TDP than the 3090, and the 4080 is the same as 3080, but everything else is lower. So I'm not really sure what you're referring to here.

The design costs for these chips is starting to starting to stretch into the $Billions. Also, people tend to be dismissive of software costs. Something like 90% of Nvidia's engineers are software. They need to get paid, too! Traditionally, you can ignore these NRE costs, but I think they're actually big enough that I think you simply cannot just look at chip fabrication costs, to establish a cost basis for them.

Unless there's large amount of hiring, or amazing raises, these costs do not justify the type of price increase coupled with lower performance increase generation on generation (aside from 4090). The 40 series smacks of nvidia pulling an apple to maintain profits with down sales (the first year the iphone was down in sales apple made more money than the prior year due to price increases). The 40 series architecture it absolutely fantastic, but the pricing leaves a lot to be desired at most tiers and shouldn't be defended.

 

[bit_user]

The 4070 uses the same die as the 4070ti and the cost to manufacture is die size so it doesn't matter how much is "enabled". The yield might be slightly higher, but there's a $200 MSRP difference between the two.

Basically, how it works is they take each die and test it. Those which don't clock high enough or have enough functioning CUDA cores go in a lower bin and are sold as 4070. At some point, that bin might run empty, at which point they have to use some higher-quality dies that could've been sold as 4070 Ti, but that also probably means their yields have improved on the node and they probably have enough top-binned dies to spare.

The reason I'm comparing the RTX 4070 Ti with the RTX 4090 is that they both launched reasonably close in time. So, we can presume their yields were similar. Their pricing must have been based on those yields.

They don't though, because they're largely a lot smaller than their 30 series counterparts

Not really. It just seems that way, because the 3000-series had 4 die sizes and the 4000-series has 5. At certain points, you get a mismatch which makes it seem like the 4000 die is smaller, but the actual dies are roughly the same size. So, the thing to do is set aside product naming and just look at die size & price.

Ampere Die	Ampere Area	Ampere Launch MSRP	Ada Die	Ada Area	Ada Launch MSRP
GA102	628​	$1,500​	AD102	609​	$1,600​
GA104	393​	$500​	AD103	379​	$1,200​
GA106	276​	$330​	AD104	295​	$800​
GA107	200​	$250​	AD106	188​	$400​
			AD107	159​	$300​

The outlier is really the AD107, which also is weird because it's 85% as big as its elder sibling. It's almost hard to believe they bothered with it - must've been some other reason.

I'm talking about the SKUs and subsequent pricing which we all watched them change in public after backlash (12GB 4080 turning into 4070ti).

That's one case where they overreached with their pricing, but they could afford to back off because the AD104 is already 2 steps down from the AD102. If I were more cynical, I'd suggest it was a manufactured controversy to make them seem reasonable and flexible.

It's pretty obvious that the 4060ti-4070ti are all 1 tier off what they really should have been when comparing performance level versus prior generation uplifts.

The GPU price didn't double the wafer price did which means the prices went up, but how much is dependent entirely on how many chips per wafer they're getting. The 4060ti will forever be my punching bag example here because it's less than half the size of the 3060ti.

I say don't get hung up on names. Just look at perf/$ (or perf/W, GB/$, etc.) and look at the ones which are out of line.

Unless there's large amount of hiring, or amazing raises, these costs do not justify the type of price increase coupled with lower performance increase generation on generation (aside from 4090).

According to this, design costs are increasing very rapidly:

Synopsys Intros AI-Powered EDA Suite to Accelerate Chip Design and Cut Costs

www.anandtech.com

On top of that, tech salaries have risen a lot in the past 3 years, which compounds with any hiring they've done. Mainly, I'm saying everyone wants to treat NRE costs as $0, when they're most definitely not.

The 40 series smacks of nvidia pulling an apple to maintain profits with down sales (the first year the iphone was down in sales apple made more money than the prior year due to price increases). The 40 series architecture it absolutely fantastic, but the pricing leaves a lot to be desired at most tiers and shouldn't be defended.

I'm just saying that the RTX 4070 Ti's price is as defensible as the RTX 4090. They're equally overpriced or not, based on die area and memory capacity. The RTX 4080 should be $150 cheaper, in order to bring it in line with those two. So, it's the most overpriced of the 3. That's all I'm saying.

 

[thestryker]

Basically, how it works is they take each die and test it. Those which don't clock high enough or have enough functioning CUDA cores go in a lower bin and are sold as 4070. At some point, that bin might run empty, at which point they have to use some higher-quality dies that could've been sold as 4070 Ti, but that also probably means their yields have improved on the node and they probably have enough top-binned dies to spare.

The cost for both is the exact same amount because they're the same size. The only difference is that more chips are going to be able to be used as 4070s than 4070ti.

The reason I'm comparing the RTX 4070 Ti with the RTX 4090 is that they both launched reasonably close in time. So, we can presume their yields were similar. Their pricing must have been based on those yields.

They've undoubtedly been stockpiling die for use in the 4070 since day one of producing the AD104 die. .

Not really. It just seems that way, because the 3000-series had 4 die sizes and the 4000-series has 5. At certain points, you get a mismatch which makes it seem like the 4000 die is smaller, but the actual dies are roughly the same size. So, the thing to do is set aside product naming and just look at die size & price.

​

​

Ampere Die	Ampere Area	Ampere Launch MSRP	Ada Die	Ada Area	Ada Launch MSRP
GA102	628​	$1,500​	AD102	609​	$1,600​
GA104	393​	$500​	AD103	379​	$1,200​
GA106	276​	$330​	AD104	295​	$800​
GA107	200​	$250​	AD106	188​	$400​
			AD107	159​	$300​

You cannot disconnect product name from die size if you're trying to use product MSRP. GA102 launched in $700/$1500 products and then die harvested into (no msrp)/$1200/$2000 products. None of this changes the cost of the die itself it's just maximizing return.

That's one case where they overreached with their pricing, but they could afford to back off because the AD104 is already 2 steps down from the AD102. If I were more cynical, I'd suggest it was a manufactured controversy to make them seem reasonable and flexible..

AD103 costs maybe 10% more per die than GA102. The 4080 has more memory, but uses 2 fewer chips with its less wide memory bus and has the same power requirements as the 3080. In what way is going from $700 to $1200 not an overreach in pricing here?

I say don't get hung up on names. Just look at perf/$ (or perf/W, GB/$, etc.) and look at the ones which are out of line.

Edit: apparently posted on accident without finishing.

If you are talking cost to manufacture you have to keep in mind everything. When you compare to prior generations this is the smallest uplift gen over gen since at least the 700 series cards (ignoring the 4090) when you compare on price/SKU.

 

[bit_user]

The cost for both is the exact same amount because they're the same size. The only difference is that more chips are going to be able to be used as 4070s than 4070ti.

Depending on yield, your cost for the Ti might be higher, because you might need to fab more wafers than you'd otherwise need, just to get enough of them.

You cannot disconnect product name from die size

I just did. You should try it, sometime. Compute the perf/$ of each product. Cover up the names and just look at which ones are out of line.

GA102 launched in $700/$1500 products and then die harvested into

That's why I looked at the MSRP of the product which launched with a given die. That reflects the truest price of the die.

The 4080 has more memory, but uses 2 fewer chips with its less wide memory bus and has the same power requirements as the 3080. In what way is going from $700 to $1200 not an overreach in pricing here?

You're getting hung up on names. I keep saying: don't do that.

 

[thestryker]

According to this, design costs are increasing very rapidly:

​

Synopsys Intros AI-Powered EDA Suite to Accelerate Chip Design and Cut Costs

www.anandtech.com

​

On top of that, tech salaries have risen a lot in the past 3 years, which compounds with any hiring they've done. Mainly, I'm saying everyone wants to treat NRE costs as $0, when they're most definitely not.

Oh they definitely are a factor without a doubt, but context is required with all of these. They didn't layoff people last year when losing money, but you can bet there weren't raises.

I'm just saying that the RTX 4070 Ti's price is as defensible as the RTX 4090. They're equally overpriced or not, based on die area and memory capacity. The RTX 4080 should be $150 cheaper, in order to bring it in line with those two. So, it's the most overpriced of the 3. That's all I'm saying.

I don't think the pricing on the halo SKU has ever been defensible, but they've always been high so it is what it is. Though from what I gathered the 3090 was an extremely expensive card to make so there wasn't much room for board partners to make profits after nvidia's margin (makes sense given the clamshell memory).

In the narrow context of the 40 series sure the pricing may make sense, but the 40 series doesn't exist in a vacuum.

 

[thestryker]

Depending on yield, your cost for the Ti might be higher, because you might need to fab more wafers than you'd otherwise need, just to get enough of them.

The Ti undoubtedly costs more because fewer chips will pass muster, but none of the usable chips are going to waste.

I just did. You should try it, sometime. Compute the perf/$ of each product. Cover up the names and just look at which ones are out of line.

I'll use data from a prior post I made as I don't really feel like grabbing more so which one of these stands out:
Launch MSRP (approx CPI inflation):
GTX 770 $399 ($520)
GTX 970 $329 ($420)
GTX 1070 $379/449 FE ($480/565)
RTX 2070 $499 ($600)
RTX 3070 $499 ($585)
RTX 4070 $599

just going by some TPU 1080p numbers:
GTX 770 26% slower than GTX 970
GTX 970 36% slower than GTX 1070
GTX 1070 25% slower than RTX 2070
RTX 2070 28% slower than RTX 3070
RTX 3070 17% slower than RTX 4070

That's why I looked at the MSRP of the product which launched with a given die. That reflects the truest price of the die.

I know which is why I did the same. Perhaps you don't remember that the 3080 and 3090 were launched at the same time.

You're getting hung up on names. I keep saying: don't do that.

Okay I paid a little over $800 for my 12GB 3080 for the same price I could get a 4070ti which will get roughly 16% performance increase at 1440p. My card is roughly 29% faster than the RTX 2080ti at the same resolution.

 

[bit_user]

I'll use data from a prior post I made as I don't really feel like grabbing more so which one of these stands out:
Launch MSRP (approx CPI inflation):
GTX 770 $399 ($520)
GTX 970 $329 ($420)
GTX 1070 $379/449 FE ($480/565)
RTX 2070 $499 ($600)
RTX 3070 $499 ($585)
RTX 4070 $599

just going by some TPU 1080p numbers:
GTX 770 26% slower than GTX 970
GTX 970 36% slower than GTX 1070
GTX 1070 25% slower than RTX 2070
RTX 2070 28% slower than RTX 3070
RTX 3070 17% slower than RTX 4070

That's the opposite of what I said. Ignore the names and just look at performance and cost.

Nvidia only introduced the 90-series in the RTX 3000 generation. That says they're still monkeying with their product tiers and maybe hadn't finished yet. People are getting tied in knots over these names, but ultimately names don't necessarily mean anything. So, just ignore them.

Okay I paid a little over $800 for my 12GB 3080

New? When?

 

[thestryker]

That's the opposite of what I said. Ignore the names and just look at performance and cost.

That entire list is about nothing but performance and price if you can't see that I've got nothing else for you on this topic.

New? When?

Brand new immediately the first time GPU prices dropped back to earth as I was watching prices regularly.

But you can always compared the regular $700 3080 where the performance boost of a 4070ti is about 21% with the 3080 about 23% faster than the 2080ti.

 

[bit_user]

That entire list is about nothing but performance and price if you can't see that I've got nothing else for you on this topic.

You're comparing model numbers across generations. Stop doing that. It's not helpful.

The only thing you should be looking at is perf/$ (and GB/$, I suppose). The names are a complete distraction.

 

[thestryker]

You're comparing model numbers across generations. Stop doing that. It's not helpful.

The only thing you should be looking at is perf/$ (and GB/$, I suppose). The names are a complete distraction.

This will be my last post on this specific bit:

You're the one getting stuck on model names that post lists the performance along with the price. You seem incapable of seeing beyond what you want to so I'll adjust for inflation just to make it even simpler. If ~$600 (MSRP) in 2023 money buys you 25% more performance (2070) then 28% more performance next gen (3070) followed by 17% more performance (4070) that last one is giving you less for your money.

I don't know why you're so set on trying to stick up for the pricing of the 40 series when it's just not good. It's a price reset like the 20 series was, but they're giving you even less uplift for your money.

edit: perf/$ only makes sense when deciding what to buy today not historical as it always gets better every generation it's just a matter of how much.

 

[bit_user]

You're the one getting stuck on model names that post lists the performance along with the price. You seem incapable of seeing beyond what you want to

It lists performance relative to the same model name of the previous generation - not pure perf/$. You're incapable of seeing beyond what you want to look at.

If you want to get all wrapped up in model names, that's your problem. It's a false problem, though. The only thing that matters is perf/$ and GB/$ Everything else is a distraction.

I don't know why you're so set on trying to stick up for the pricing of the 40 series

I never said it was good. The only thing I said was that their costs have gone up, giving them less room to cut prices than I think most people assume. However, the pricing decisions were pretty much baked into their choice of die sizes and process node. Once those were finalized, the main variable they had left to alter was how to name them.

 

[Gescom]

Let's talk ratios, not increases. So, 5 nm prices are 283.5% as high as 10 nm. With density being 329.5% as high, then the cost per 5 nm transistor is 86.1% as high.

Next, let's look at the relative transistor counts for AD102 vs. GA102. That's 76.3 / 28.3 = 269.6% as many. So, with the per-Tr savings of 5nm, you'd expect the GPU chip in the RTX 4090 to cost 232.0% as much.

And that's probably an underestimate, since AD102 has so much more L2 cache, which we know is less dense than logic. I think it's misleading to compute this on the basis of per-transistor costs, when those have an in-built assumption about the ratio of different types of cells. It's better to simply compute them on an area-basis.

BTW, Nvidia didn't even use TSMC for GA102, so we really should be comparing Samsung 8 nm. And they didn't use TSMC N5 for AD102, but rather 4N (in lieu of any specifics, we can treat it as N5 but it's not).

Getting back to TSMC, it would be illuminating to compare the per-Tr cost reductions of 7nm -> 5nm with earlier node transitions. I'm sure you'll find their per-Tr costs reduced far less this time than in the past.

Agreed, my wafer price estimations are probably way off. Tsmc's 5nm could be $20k per wafer but also $12k for an early volume customer.


TSMC is largely responsible for NVIDIA prices which was the initial claim.

People are complaining about GPU prices, but from all the reports I have read, TSMC is largely responsible for these price hikes, much more than Nvidia or AMD themselves.

rtx 3060 TI / ~1.3x 2060S perf
2020, $399
Transistors 17.4m
Density 44.4m / mm²

rtx 4060 TI / ~1.1x 3060 TI perf
2022, $399
Transistors 22.9m
Density 120.5m / mm²
--
rtx 2080 / ~ 1.4x 1080 perf
2018, $699
Transistors 13.6m
Density 25m / mm²

rtx 3080 / ~ 1.6x 2080 perf
2020, $699
Transistors 28.3m
Density 45.1m / mm²

rtx 4080 / ~1.5x 3080 perf
2022, $1199
Transistors 45.9m
Density 121.1m / mm²
--
rtx 3090
2020, $1499
Transistors 28.3m
Density 45.1m / mm²

rtx 4090 / ~ 1.6x 3090 perf
2022, $1599
Transistors 76.3m
Density 125.5m / mm²
--
Comparing prices vs transistors/density vs perf between NVIDIA's same gen gpus and segments, it is obvious that TSMC isn't largely responsible for Nvidia's price hike. NVIDIA is and yes, it's a mess.

 

[TJ Hooker]

TSMC has been aggressively raising prices on high-end nodes in the last 3 years. Far higher than inflationary pressure can possibly explain.

The reason these price hikes are so extremely aggressive on high-end nodes but less on low-end, is a reflection of TSMC competition on low-end nodes, but monopoly on high-end nodes

Your figure seems to show the exact opposite of what you're saying. The older, low-end nodes have seen larger relative price increases over the last few years, compared to the newer, high-end nodes. The newest node (5nm) shown has decreased in price, the only node to do so.

 

[bit_user]

Your figure seems to show the exact opposite of what you're saying. The older, low-end nodes have seen larger relative price increases over the last few years, compared to the newer, high-end nodes. The newest node (5nm) shown has decreased in price, the only node to do so.

I think I can explain.

You normally have a sharp reduction in its cost, immediately following the introduction of a new node. Eventually, it levels out.

The nodes besides 5 nm are all older nodes that have been around for longer than the 3-year period from which the data was taken. Because they're older, mature nodes, you'd expect their prices to be stable, yet they increase.

When this same "inflation" effect is applied to the much newer 5 nm node, it's working against the normal new-node cost-reduction curve. The result is a premature leveling-out of its cost, as the inflation rate between 2021-2022 counteracts the cost reductions they achieved in that time.

As for why the oldest nodes increased the most? I don't suppose we can tell whether that's more due to market conditions or that basically every wafer you make probably has some fairly fixed costs and any increases in those are most visible on the cheapest nodes. I guess another possibility is that they managed to achieve cost reductions on some of the newer nodes to partially offset the increases.